Live Feed
First-Ever Three-Way Tie at Intelligence Index 57: All Three Frontier Labs Now Equal
Artificial Analysis has called a tie at the top of its Intelligence Index for the first time. Claude Opus 4.7 (57.3), Gemini 3.1 Pro Preview (57.2), and GPT-5.4 (56.8) all round to 57 within a 1-point confidence interval. The composite ranking is dead. Lab differentiation is the new story.
Special Report: The AI Index 2026 — Signals From a Frontier That Refuses to Slow Down
The ninth AI Index puts hard numbers on a year the frontier kept accelerating, the US–China gap effectively closed, private capital bent toward the United States, and the hardware stack concentrated around a single foundry. A data-forward read of where the model race actually is.
Anthropic Puts Claude Inside Microsoft Word With Native Tracked Changes — Legal Review Is the Opening Move
Anthropic has released a beta add-in placing Claude directly inside Microsoft Word. Every AI-generated edit surfaces as a native tracked change — not a sidebar suggestion. Legal contract review is the lead use case, signalling Anthropic is going after the enterprise document workflow that anchors Word's commercial moat.
llama.cpp Merges Speculative Checkpointing: 40% Less VRAM, 20% More Throughput for Consumer 70B Inference
Georgi Gerganov merged speculative checkpointing into llama.cpp on April 18, cutting VRAM usage by up to 40% during batched operations and boosting token throughput by 15–20% on consumer hardware. The update makes extended-context 70B inference viable on setups that previously ran out of memory.
Opus 4.7 Opens a 79-Point Agentic Elo Lead Over GPT-5.4 — and Runs Cheaper Per Task Than Its Predecessor
Claude Opus 4.7 scores 1,753 on GDPval-AA, 79 Elo points ahead of Sonnet 4.6 and GPT-5.4 (1,674) and 134 ahead of Opus 4.6 (1,619). Despite identical list pricing at $5/$25 per million tokens, effective per-task cost is lower than Opus 4.6 Adaptive Reasoning — driven by reduced output token usage in Opus 4.7's new tokenizer.
Opus 4.7's Hidden Cost: New Tokenizer Adds Up to 35% More Tokens Per Prompt
Anthropic held the line at $5/$25 per million tokens for Claude Opus 4.7, but confirmed the new model uses up to 35% more tokens for the same input. Developer reports also flag a collapse in long-context retrieval performance — from 78.3% to 32.2% on MRCR.
xAI Opens Grok Voice APIs: $0.10/hr Transcription, $4.20/M Characters Speech Synthesis
xAI has launched standalone Grok Speech-to-Text and Text-to-Speech APIs built on the same production stack powering Grok mobile, Tesla vehicles, and Starlink customer support. The release positions xAI directly against ElevenLabs, Deepgram, and AssemblyAI with sharp pricing and a proven infra pedigree.
Anthropic Launches Claude Design: Conversational Prototyping That Sent Figma Down 7%
Anthropic shipped Claude Design in research preview on April 17, putting prototypes, wireframes, pitch decks, and marketing assets behind a text prompt. Figma shares dropped 6.84% on the announcement. The tool runs on Opus 4.7, exports to Canva/PDF/PPTX/HTML, and hands off directly to Claude Code.
Cerebras Files for Nasdaq IPO With $20B OpenAI Contract and $510M in Revenue
AI chipmaker Cerebras Systems disclosed its S-1 on April 17, revealing $510M in 2025 revenue, a $20B multi-year OpenAI deployment deal, and an AWS data center integration. The listing breaks an 18-month freeze on AI hardware IPOs.
Cloudflare Ships Five Agent Products in Three Days — Project Think, Mesh, and the Bid to Own Agent Infrastructure
Cloudflare ran "Agents Week" April 15–17, launching Project Think (a batteries-included Agents SDK with durable execution and sub-agents), Cloudflare Mesh (private networking for agent fleets), Agent Lee (dashboard AI assistant), Redirects for AI Training, and an Agent Readiness Score. It is the most concentrated infrastructure land-grab since the cloud wars.
Meta to Cut 8,000 Jobs Starting May 20 — The Largest AI-Driven Headcount-to-Compute Swap in Tech History
Meta will begin laying off roughly 10% of its global workforce on May 20, 2026, with additional rounds expected later in the year. The capital is being redirected into chips, data centres, and model training — making this the single largest AI-driven reallocation of labour to compute in tech history.
MiniMax Open-Sources M2.7 at 1495 GDPval-AA ELO and 57% Terminal Bench — Self-Evolving Agent Hits Top-5
MiniMax has released M2.7 under an open-source license, posting 1495 ELO on GDPval-AA (highest at release), 56.22% on SWE-Pro, and 57.0% on Terminal Bench 2. A Chinese lab has put a freely self-hostable model inside the global top tier on agentic capability benchmarks.
MIT, Oxford, CMU Study: 10 Minutes of AI Assistance Measurably Reduces Independent Problem-Solving Persistence
A multi-institution study across 1,200 participants finds that direct AI assistance improves immediate performance but reduces problem-solving persistence within minutes of the AI being removed. The damage is locatable: it comes from answer-outsourcing, not AI use itself.
Bank of England, Fed, and Treasury Convene Emergency Sessions Over Claude Mythos — White House Extends Access to Federal Agencies
Finance ministers and central bank governors worldwide are holding crisis meetings after Claude Mythos demonstrated the ability to chain previously unknown vulnerabilities across banking and critical infrastructure software. The White House is now planning Mythos access for major federal agencies.
OpenAI Codex Becomes a macOS Desktop Agent — Computer Use, Memory, and 90+ Plugins Ship Together
OpenAI has transformed Codex from a coding assistant into a persistent macOS agent that can see, click, browse, generate images, and remember user habits across sessions. The update is a direct escalation against Anthropic's Claude Code and positions Codex for OpenAI's planned super-app.